MIMIC / fox /metadata.json
erickfm's picture
Upload MIMIC character checkpoints + model card
f5d1cd7 verified
raw
history blame contribute delete
784 Bytes
{
"character": "Fox",
"melee_enum": "FOX",
"run_name": "fox-20260414-relpos",
"global_step": 55692,
"games_trained": 17319,
"val_btn_f1": "87.1%",
"val_main_f1": "~55%",
"val_loss": "0.77",
"n_params": 26276389,
"n_controller_combos": 7,
"model_preset": "mimic",
"pos_enc": "relpos",
"no_self_inputs": false,
"training_notes": "Trained 2026-04-14, fox-20260414-relpos run. v2 shards, Shaw relative-position attention (mimic preset), --self-inputs, dropout 0.2, 65K steps at batch 512 (2x the standard schedule). Replaces the yesterday's RoPE Fox, which matched this on val metrics but relpos won the 2026-04-14 sweep on Falco so we standardized on it for all new runs. Best checkpoint is the step 55692 snapshot, closest to the lowest measured val loss."
}